Robust and generalized functional data approach for yield curve modelling of the Asian frontier government bond market for steady state and non-steady state

Abstract

Studies in Yield curve modelling have a long history in the financial capital market. Its ability to produce substantial insights about the future behaviour of interest rates and economic performance has encouraged many scholars to refine further the model’s suite for complex, dynamic global capital markets. The state-of-the-art yield curve model structure developed by Nelson-Siegel is still being used by many monetary authorities. However, the accuracy of the model has been questioned theoretically as well as its applicability, especially to frontier markets, due to the inherent limitations of the model. Despite many studies having been done to improve the Nelson-Siegel model to capture the dynamic nature in the developed capital market, while addressing the weaknesses, limited studies have focused on the frontier markets. Due to the lack of sequential data availability and not enough depth of the tenor vis data, three- dimensional yield curve predictions are left in a dilemma for countries like Sri Lanka. The study examined the yield curve data of Sri Lanka, Bangladesh and Pakistan while deriving the zero-coupon bond yield from the discounted yields. Non-traditional smoothing tools such as super-smoothing and spline smoothing have been found to improve the accuracy of the Nelson-Siegel model. The functional principal component analysis suggested the existence of a strong functional structure. However, the study examined the multiple univariate approach of the yield curve prediction using deep learning architectures such as LSTM, RNN, CNN and TCN. Mean Square Error indicated that the multivariate approach exhibits higher accuracy as the yields of each tenor collectively influence the next step of the yield matrix. Utility of the Gompertz function supersedes the Nelson-Siegel model in the Sri Lankan and Bangaldesh context, while the multivariate LSTM architecture outperforms the rest in the Pakistan yield curve modelling after hyperparameter tuning.